Multimedia Tools and Applications

Multimedia Tools and Applications SCM-motivated Enhanced CV Model for Mass segmentation from Coarse-to-Fine in Digital Mammography --Manuscript Draft-Manuscript Number:

MTAP-D-17-01261R1

Full Title:

SCM-motivated Enhanced CV Model for Mass segmentation from Coarse-to-Fine in Digital Mammography

Article Type:

Manuscript

Keywords:

Mammography, Mass Segmentation, Spiking Cortical Model (SCM), Improved CV model, Local Region-Scalable Force (LRSF).

Corresponding Author:

Yide Ma Lanzhou University Lanzhou, Gansu CHINA

Corresponding Author Secondary Information: Corresponding Author's Institution:

Lanzhou University

Corresponding Author's Secondary Institution: First Author:

Ya'nan Guo

First Author Secondary Information: Order of Authors:

Ya'nan Guo Xiaoli Gao Zhen Yang Jing Lian Shiqiang Du Huaiqing Zhang Yide Ma

Order of Authors Secondary Information: Funding Information:

Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation

Response to Reviewer Comments

Dear Editors and Reviewers: Thanks for your letter and the reviewer’s comments about this manuscript entitled “SCM-motivated Enhanced CV for Mass segmentation from Coarse-to-Fine in Digital Mammography” (No. MTAP-D17-01261). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our research. We have studied these comments carefully and have made corrections accordingly. In revised manuscript, figures are renumbered to make the article more clearly. Meanwhile, in order to show the difference between the original manuscript and the revised, as suggested, we list the major change as below and structure of the whole manuscript are not listed here.

Response to Reviewer #1: Reviewer #1: Thank the authors for the submission of the manuscript. The paper is well-organized and compact: a) the abstract is informative and contains the exact information so as the reader gets a clear idea on the entire manuscript, b) the technical analysis seems accurate, c) qualitative, quantitative and comparative results are included for the validation of the proposed scheme. In order to further improve the quality of the paper, the following amendments are proposed: Comment 1: the paper is based on a previous work of the authors, please indicate in short, what the new characteristics, considerations, improvements are with respect to the initial implementation, since a combination of the Spiking Cortical Model (SCM) and the Improved CV Model has already been adopted. This will also clarify what the novelty of the current work is. Thanks for this comment. This paper is an extended version of our preliminary work presented in a conference paper. In previous work, we proposed the combined method of SCM-motivated Improved CV Model based on bias field, although it results in great improvement in accuracy for mass detection, there exists obvious distortions in obtained results, especially for mass mammograms where the contrast is relatively low and boundary is blurry, the obtained contour cannot completely converges to the mass boundary. Thus, we introduce local region-scalable external force (LRSF) to enhance above model to obtain a sophisticated segmentation method. In former manuscript, we cannot give the clear and detailed description about proposed model, so in revised paper, we first give the corrected cite of previous work, and then provide clear explanation about the goal of our work in line 7-27 of page 2, the enhanced theory about proposed SCM-motivated enhanced CV model can be seen in subsection “C. Mass

Candidate

Localization ” and subsection “D.

Segmentation Methods”, and thank you very much for your

reminder, these changes let the theory of our method more clear and readable. Comment 2: analysis of related work is short. In addition, the limitations of the state-of-the-art approaches (apart from the active-contours based ones) the proposed work intends to overcome are not clear to the reader. Please enrich the introduction section and include a short paragraph emphasizing on the above mentioned feature. Thanks for this comment. As you said, the related work is too short, so we enriched the introduction in revised version, further explanation can be found in line 7-16 of page 2; furthermore, we have added some classical and boosting techniques to improve and perfect my work in “Introduction” section. Here, we list them in table.1. table.1 Added references Added references Initialization Techniques for Segmentation with the Chan-Vese Model

Ref [19]

An Optimal Initialization Technique for Improving the Segmentation Performance of Chan-Vese Model

Ref [20]

Distance regularized level set evolution and its application to image segmentation

Ref [21]

Automated geographic atrophy segmentation for SD-OCT images using region-based C-V model via local similarity factor

Ref [22]

Efficient segmentation of piecewise smooth images

Ref [23]

Localizing region-based active contours

Ref [24]

Coefficient of Variation Based Image Selective Segmentation Model Using Active Contours

Ref [25]

A variational model with hybrid images data fitting energies for segmentation of images with intensity inhomogeneity

Ref [26]

Fast Global Minimization of the Active Contour/Snake Model

Ref [27]

An efficient local Chan-Vese model for image segmentation

Ref [28]

A level set method for image segmentation in the presence of intensity inhomogeneities with application to MRI

Ref [31]

Comment 3: in equation (1), please explain notations kl and KL. Also, in equation (8) describe quantities M, L and R and in equations (9), (10) indicate the difference between Hε and H, utilized in equation (4). Thanks for this comment. Your reminder let us find some mistakes in former version, kl and KL has the same meaning in equation (1), we replace the KL using kl, (k,l) denotes the location of adjacent neurons. In order to give more clear explanation of proposed model, we use the following equation (a) replace the equation (8) in former paper:





F CV  , c1, c2      K  y  x  I  x   b  y  ci dyM i   x   dx  V  H   x  dx N

2



i 1

Here,



M

i

denotes membership functions, For two-phase case, N







2,

(a)

so its membership functions

are M1   H  and M 2   1  H  , H () is the Heaviside function. More detailed description

can be seen in the second-to-last paragraph of page 5. As a matter of fact, Hε and H have the same meaning in this work, so we replace all the Hε using H in revised paper, and H denotes the Heaviside function. Comment 4: for the reproducibility of comparative results, please give information and/or references to describe the region-based model with bias field (Method 1) and typical CV model (Method 2). Thanks for this comment, we provide the correct references to Method 1 and Method 2, which can be found in line 25 of page 2. Method 1 is a region-based model with bias field, and Method 2 denotes a typical CV model. Comment 5: for better discrimination and understanding, in Fig. 4 please include a short label on the left or right side of each row (e.g. proposed method, proposed method enlarged). The same can be done for the results depicted in Fig. 5. Thanks for your comment, according to your advice, enriched description is added in Fig. 4 and Fig. 5 for better discrimination and understanding, and thank you again. Comment 6: regarding the description of results depicted in Fig. 5, please indicate the difference between "ground truth segmented manually" and "ground truth provided by expertise" (who performed the manual segmentation and why is this dual information used?). Thanks for this comment, in MIAS database, the experts just provide the ground truth that is located by the center coordinates and the radius, which is a circle region and is very different to its real edge of the lump, the aim of proposed model is to dig the fine mass segmentation results, so we regard the manual segmentation that is very close to real mass areas as ground truth to measure the segmentation ability of the method. Here, we provide the ground truth labeled by red circle as a reference, from Fig.5-(e-4)-(h4) in revised paper, we also can find the great difference between ground truth and manual segmentation, more detailed information can refer to [1], and we also provide enhanced explanation in line 3-5 of paragraph 1 in page 9 for revised manuscript. Comment 7: in Table 1, comparative results for Method 1 and Method 2 can be included too, so as to indicate any significant differentiation and enhance validated outperformance of proposed scheme. Also, clarify if the whole reference databases have been utilized too for extracting the results depicted in Table 2. Thanks for your comment, although the detection rate, mean and variance of the area overlap metric can make a certain assessment of the algorithm, it is a rough evaluation, and the more detailed evaluation

should be based on average sensitivity, specificity and Dice Similarity Co-efficient, even for ROC curve. More widely-accepted measurement are more persuasive than others, so here, we don’t list the comparative results for Method 1 and Method 2 after the comprehensive consideration of your advices, and thank you for your reminder. Comment 8: comparative results of Tables 2-6 could contain the average performance for all images in the datasets. Instead, the criterion for the selection of the specific 8 images should be explained (e.g. why these 8 reference images can represent the entire database?). Thanks for your comment. In DDSM database, we list four images to set the example, image (a) and (b) are benign mammograms, and image (c) and (d) are malignant mammograms. In MIAS database,

image (e) and (f) are benign, (g) and (h) are malignant. All these 8 images can give the representative of all the mass in these two databases, and proposed method can achieve excellent mass segmentation not only for benign image but also for malignant one. Comment 9: for the construction of the ROC please explain the procedure followed (how many images were utilized, how many operating points were selected, what thresholds were used for each point?). Thanks for this comment, receiver operating characteristic curve (ROC) can be plotted by average the sensitivity and specificity. Typically, the sensitivity is regarded as Y-axis, representing the true positive rate, 1-specificity is regarded as X-axis representing the false positive rate, and then we can easily obtain the ROC curve by the value of sensitivity and 1-specificity. Comment 10: please include authors' affiliations in the beginning of the manuscript. Thanks for this comment, we have already added the authors' affiliations in revised manuscript. Comment 11: references 15-17 to represent "recent work on active contour implemented via level set methods to address a wide range of image segmentation problems in image processing and computer vision" should be more recent. Please update them with new, representative ones. Thanks for this comment, we have already replaced some of them using the latest ones, and the added literates is listed in table 2. table 2 Added references Added references Automated Robust Image Segmentation: Level Set Method Using Nonnegative Matrix Factorization with Application

Ref [9]

2016

An Optimal Initialization Technique for Improving theMRI Segmentation Performance of Chan-Vese Model to Brain

Ref [20]

2017

Comment 12: reference [11] does not contain the complete information. Thanks for your reminder, we have provided the complete information in revised version.

Comment 13: the manuscript needs a deep check for proofreading. Thanks for this comment, we have made three time check for proofreading, and the improved edition has higher readability than before, and thank you again. Reference [1] R. Lacson, K. Harris, P. Brawarsky, T.D. Tosteson, T. Onega, A.N.A. Tosteson, A. Kaye, I. Gonzalez, R. Birdwell, J.S. Haas, Evaluation of an Automated Information Extraction Tool for Imaging Data Elements to Populate a Breast Cancer Screening Registry, Journal of Digital Imaging, 28 (2015) 1-9.

Response to Reviewer #2: Comment: The authors have proposed a new methodology to sense unexpected mass in mammography medical images. An improvement to OpenCV's implementation is very well declared with segmentation and localization. In 4th section, improved CV Model, the steps for detection of anomaly even on the weak contour is given. From a general point of view, the paper has been presented in a well manner. Comparisons and error of the methodology is also declared. I suggest that the paper might be accepted in its current form. Thank you for your approval about our work, and we will continue to make great progress in this field.

Response to Reviewer #3: In this paper, the authors proposed a new SCM-motivated improved CV model for mass fine segmentation. This paper is well written and easy to follow. I would like to accept this paper if my following concerns are carefully addressed. Comment 1: Although the authors have conducted a thorough literature review, the following references are still missing. I would like the authors to cite the following references in the revision. Semantic Pooling for Complex Event Analysis in Untrimmed Videos, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 8, pp. 1617-1632, 2017. Semi-supervised Feature Analysis by Mining Correlations among Multiple Tasks, IEEE Transactions on Neural Networks and Learning Systems, vol. 28, no. 10, pp. 2294-2305, 2017.

Beyond Trace Ratio: Weighted Harmonic Mean of Trace Ratios for Multiclass Discriminant Analysis. IEEE Transactions on Knowledge and Data Engineering, Vol. 29, No. 10, pp. 2100-2110, 2017. Thank you for your reminder, and those updated references are added in conclusion section for future prospect, you can find them according to following index in table.3: table.3 Added references Added references Semantic Pooling for Complex Event Analysis in Untrimmed Videos

Ref [44]

Semi-supervised Feature Analysis by Mining Correlations among Multiple Tasks

Ref [45]

Beyond Trace Ratio: Weighted Harmonic Mean of Trace Ratios for Multiclass Discriminant Analysis

Ref [46]

Comment 2: The computational complexity of the proposed model should be discussed in the revision. Thank you for this comment, in this work, our main concern is to achieve fine mass segmentation results based on enhanced CV model, so we don’t pay more attention to computational complexity, in fact, proposed enhanced CV model can achieve excellent segmentation result after 50 iterations, on the contrary, the traditional model needs iterative 200 times or more to get a relatively good segmentation results. In our future work, we can give the consideration of computational complexity according to your suggestion, and thank you in advance. Comment 3: Please discuss in more detail the future direction of this work. Thank you for your reminder, inspired by the references you provided, we discovery our model can be recommended in other field such as complex event analysis, feature analysis and visual tracking, all these statement can be found in paragraph 2 of section “IV. CONCLUSIONS”. Comment 3: Based on the above comments, I would like to accept this paper if my concerns are carefully addressed. Thank you for your comments, above comments are already carefully addressed, and all these changes improve the overall level of our article, and thank you again.

Response to Reviewer #4: Comment: Well written paper with clear contribution Thank you for your affirmation of our work, and we will make persistent efforts in this field.

Response to Reviewer #5: This paper proposes a SCM-motivated improved CV model for mass segmentation from digital mammography. By exploiting the Spiking Cortical Model for mammary-specific and mass edge detection, the lesion can be located delicately, which enhances the insensibility of the improved CV model to the initial contour. By combining the principle of physical imaging, bias field and local region scalable force, the improved CV model can obtain fine mass boundary accurately. Comparison experiment on two wellknown datasets shows that the proposed method performs better than other two methods. This work is strongly motivated and the algorithm is technically sound. The performance seems much better than other two compared methods. I think this work is appropriate for publication if the following issues can be well addressed: Comment 1: Some steps or formulations of the algorithm is not clearly explained. It is hard to interpret why (6) can be transformed from (5). How is (6) related to the assumptions (A1) and (A2). The definition of c_i is not clear. How to determine the parameter \pho? I suggest the authors should give more details and explain why it should be done in this way ( e.g. setting of parameters)? Thank you for your reminder, what you said make us realize that the previous theoretical information is a little confusing, so we provide the enhanced interpretation in revised edition. You can find the great changes in subsection “Energy Formulation based on bias field”. Here, Ci denotes different constants in this work, more detailed information about can be seen in the second-to-last paragraph The original contour of enhanced CV model is obtained by SCM model, how to set the parameters of SCM model can be find in paragraph 2 of page 2. This model is presented to solve the problem of traditional CV model that poor segmentation ability when extending to segment the object with inhomogeneous intensities and the sensitiveness to initialization, especially for mass with low-contrast and blurry boundaries in mammograms, we introduced the technique of local region-scalable external force to improve the internal energy evolution mechanism in weak edge. All these explanation can be found in line 7-27 of page 2 in section “INTRODUCTION”. Comment 2: In Line 49 of Page 4, how to control the contour move to the direction with larger gray level is not clearly explained. Thank you for your reminder, we have provided the sufficient and adequate theoretical complement

in paragraph 2-3 of page 6 in revised paper according to your advice.

fig.1 Scalable local region on expandable kernel

As can be seen from Fig.1, the contour divides these windows into two parts including inside and outside regions, and we respectively conduct the greyscale statistical calculation for the inside and outside region of the contour. If statistical gray difference of the inside and outside the contour is very small or approximately equal zero in a certain window function, we believe that this window is in a homogeneous region, at this time, the contour is not the actual boundary of the object. Therefore, the pressure generated by these difference is approximately zero, so the contour line does not converge to any edge. Adjusting the size of the window to make sure both target and background region are included, and then do the statistical gray as before, in this process, we let window cover these two areas by controlling the size of l . When the gray difference between these two areas is relatively large, the evolution direction of the contour line can be judged. In mammogram, the breast mass is the brightest region in the image, which means that the target area has a larger gray area, in contrast, background has smaller ones, and the actual edge of the mass can be achieved only if the contour line evolves from higher intensities area to lower one, That is, the contour line moves to the direction with higher gray value. In the end, as long as gray level statistics difference of inside and outside region is large enough, even a very small window function is selected, our method still can converges to the true boundary breast lumps. Comment 3: Some important references are not included. In Line 39 and Line 40 of Page 2, what are the PCNN model and ICM model? From which literatures can we find these two model? In Section III, the references for Method 1 and Method 2 are not included. I suggest the author should include these important references. Thank you for your reminder, we have already added the reference that give enough explanation of PCNN and ICM model, it can be seen in ref [38]. In Section III, we also added related references on

Method 1 and Method 2, which can be seen in ref [31] and ref [33], and thank you again. Comment 4: In Section III-A, it seems that the experimental results cannot show the improved CV model has lower dependence on the initial contour. I think to show the less dependence, different initialization of contour should be used to test the algorithms. I suggest more analysis should be provided to verify the conclusion. Thank you for your reminder, as is known to all, typical CV model is sensitive to reinitialization contour, so as to achieve better segmentation results we must set the better initialization that is more close to mass real contour. In this work, we employ SCM model to automatically achieve more approximate mass contour as the initialization contour, which well solved the problem of sensitive to initial contour of CV model. As a matter of fact, the aim of our work is to handle the dilemma that reinitialization sensitivity and poor segmentation for images with intensity inhomogeneity when extending to segment the mass with low-contrast and blurry boundaries in mammograms, so we don’t pay more attention to the comparison of these works, we just aim at providing another technique to improve the CV model, broadening our horizons to find different theory to handle long-standing problems, and thank you again. Comment 5: In Section III-B, the detection rate of Method 1 and 2 is not provided. I suggest the results of Method 1 and 2 should be included for comparison. Thank you for your comment, Although the detection rate make a certain assessment of the algorithm, it is a rough evaluation, and the more detailed evaluation should be based on average sensitivity, specificity and Dice Similarity Co-efficient, even for ROC curve. More widely-accepted measurement are more persuasive than others, so here, we don’t list the comparative results of Method 1 and 2 after the comprehensive consideration of your advices, and thank you for your reminder. Comment 6: In Table 3, why does the proposed method per not perform better than Method 2 on Image a? More explanation should be given.

TP---True Positives

FP---False Positives

FN---False Negatives

Contour of ground truth

Contour of the method

fig.2 definition of TP, FP and FN.

Thank you for your comment, according to the definition of Sensitivity: SEN 

TP TP  FN

Here, we define the areas of Ground truth as GT, from the fig.2, we can see GT=TP+FN, that’s, FN=GT-TP, then SEN 

TP TP  FN



TP TP  (GT  TP )



TP

. In normal case, GT is a constant, so the Sensitivity is

GT

proportional to TP. For image a, we can see from (a-3_enlarged) for Method 2 and (a-4_enlarged) for proposed method, the TP areas of Method 2 is great larger than proposed method, so the sensitivity of Method 2 is higher than proposed method, but the Specificity of our method is almost 1.93% higher than Method 2, that’s to say, neither sensitivity nor specificity can be very objective in evaluating algorithmic performance, so in final step, AUC is introduced to measure the method combining sensitivity and specificity. Comment 7: I think the proposed algorithm have wide applications not limited to mass segmentation. It can be applied to other tasks such as object tracking in video which aims to separate the target from the background and continuously track the target. I think the author can discuss the potential application of other tasks such as object tracking and provide some related work for interest readers. I suggest the following works can be cited or reviewed. [1] Multi-cue Visual Tracking Using Robust Feature-Level Fusion Based on Joint Sparse Representation. CVPR 2014: 1194-1201 [2] Joint Sparse Representation and Robust Feature-Level Fusion for Multi-Cue Visual Tracking. IEEE Trans. Image Processing 24(12): 5826-5841 (2015) [3] Robust Joint Discriminative Feature Learning for Visual Tracking. IJCAI 2016: 3403-3410 [4] Robust visual tracking using dynamic feature weighting based on multiple dictionary learning. EUSIPCO 2016: 2166-2170 [5] Robust MIL-Based Feature Template Learning for Object Tracking. AAAI 2017: 4118-4125 Thank you for your comment, we added these references in conclusion section to give the better outlook of proposed method, you can see the statement and references in paragraph 2 of section “IV. CONCLUSIONS”, the enriched references index is displayed in table.4:

table.4 Added references Added references Multi-cue Visual Tracking Using Robust Feature-Level Fusion Based on Joint Sparse Representation

Ref [47]

Joint Sparse Representation and Robust Feature-Level Fusion for Multi-Cue Visual Tracking

Ref [48]

Robust Joint Discriminative Feature Learning for Visual Tracking

Ref [49]

Robust MIL-Based Feature Template Learning for Object Tracking

Ref [50]

Response to Reviewer #6: This paper presents an improved model for mass segmentation in digital mammography. The Spiking Cortical Model is employed for generating the initial contour of improved CV model. The experimental results show the algorithm is effective. Comment 1: There are many existing works using a coarse segmentation algorithm (such as FCM) to provide the initial contour of MS or CV model. These works need to be cited. The performance of proposed method needs to be compared to that of these works. Thank you for your comment, in revised manuscript, we gave the consideration FCM related theory to improve the CV model according to your suggestion, we also cited any other promising techniques in “INTRODUCTION” section inspired by what you said, these further explanation can be found in line 7-line16 of page 2, and added classical and boosting techniques enrich our work and improve the readability of our paper. Here, we list them in table.5. table.5 Added references Added references Initialization Techniques for Segmentation with the Chan-Vese Model

Ref [19]

An Optimal Initialization Technique for Improving the Segmentation Performance of Chan-Vese Model

Ref [20]

Distance regularized level set evolution and its application to image segmentation

Ref [21]

Automated geographic atrophy segmentation for SD-OCT images using region-based C-V model via local similarity factor

Ref [22]

Efficient segmentation of piecewise smooth images

Ref [23]

Localizing region-based active contours

Ref [24]

Coefficient of Variation Based Image Selective Segmentation Model Using Active Contours

Ref [25]

A variational model with hybrid images data fitting energies for segmentation of images with intensity inhomogeneity

Ref [26]

Fast Global Minimization of the Active Contour/Snake Model

Ref [27]

An efficient local Chan-Vese model for image segmentation

Ref [28]

A level set method for image segmentation in the presence of intensity inhomogeneities with application to MRI

Ref [31]

As a matter of fact, the aim of our work is to handle the dilemma that reinitialization sensitivity and poor segmentation for images with intensity inhomogeneity when extending to segment the mass with low-contrast and blurry boundaries in mammograms, so we don’t pay more attention to the comparison

of these works, we just aim at provide another technique to improve CV model, broadening our horizons to find different theory handle long-standing problems, and thank you again. Comment 2: Which image feature is employed as the input of SCM? How to set the parameters of SCM in this paper? These details need to be added. Thank you for your comment, in our work, intensity is considered as the input of SCM model, the parameters setting of SCM model is added in paragraph 2 of page 4, through massive tests, the parameters are set as follows:

U ij (0) initialization is set as 0, and Eij (0) initialization is set as 1, f=0.2, g=0.9, h=20,  =1.

 0.1091

Synaptic weight matrix is set as W   0.1409

  0.1091

  . Thank you for your reminder, and these  0.1091 

0.1409 0.1091 0 0.1409

0.1409

supplement make the content of our article more complete. Comment 3: The process of solving the minimization of Eq(11) need to be detailedly described. Thank you for this comment, your reminder let us find the theory of our model is very unreadable, we rewrite the theory in section “D. Segmentation Methods”, and all these change more our article more clear and straightforward. As to how to solve the minimization of energy function, many literatures provide the detailed explanation, so here, it is no longer burdensome, just providing strong reference support in Ref[17]. Comment 4: What do the symbol L and R_{p} in Eq(8) stands for？ Which window function is employed as K(y-x)? There is no descriptions in the paper. Thank you for your comment, inspired by your suggestion, we find the explanation about our model is not clear, and we rewrite the theory of our model in section “D. Segmentation Methods”. The symbol L and R_{p} are not included in revised paper, so too much explanation is not needed. K(y-x) denotes as a nonnegative window function, which is also a kernel function, here, Gaussian kernel function is applied to calculate regional statistical results. You can find the corresponding description in line 5 of paragraph 3 in page 6, and thank you very much. Comment 5: Eq(4) has made use of symbols c_{1} and c_{2}. Then Eq(6) uses symbol c_{i} again. Thank you for your comment, the symbols c_{1} and c_{2} in Eq(4) are totally different with the symbol c_{i} in Eq(6) in previous paper, so in revised edition, we changed c_{i} into confusion, thank you again for your reminder.

C

i

to avoid

Comment 6: Eq(Line 39, Page2. "PCNN" --> "Pulse Coupled Neural Network (PCNN) ". Thank you for your comment, and we have provided the full name of the abbreviation PCNN model according to your suggestion.

Response to Reviewer #8: Comment 1: I suggest that fig.1 is explained more clearly. Thank you for your comment, enhanced explanation is added in Fig.1, and thank you again. Comment 2: I suggest that section II.B is explained more clearly. Thank you for your comment, we adopt combined method to manipulate mammograms, these techniques are all mature methods, so we just provide related literatures in our work, and thank you again. Comment 3: I suggest that section III.A is explained more clearly. Thank you for your comment, in order to explain the subjective evaluation in Fig.4 and Fig.5 more clearly, we first relabeled these two figures, and then provided the enhanced explanation in subsection “A. Subjective evaluation”. Comment 4: I suggest that table 1 is explained more clearly Thank you for your comment, table 1 displays the detection rate of our model, and it just give an intuitive response to the performance of our algorithm, so we added enough explanation in sub subsection “1) Detection rate” .

Response to Reviewer #9: Comment 1: it seems that you have combined two already existing techniques SCM and CV to give your proposed technique. What is your exact contribution in designing the procedure? Thanks for this comment. This paper is an extended version of our preliminary work presented in a conference paper. In previous work, we proposed the combined method of SCM-motivated Improved CV Model based on bias field, although it results in great improvement in accuracy for mass detection, there exists obvious distortions in obtained results, especially for mass mammograms where the contrast is relatively low and boundary is blurry, the obtained contour cannot completely converges to the mass boundary. Thus, we introduce local region-scalable external force (LRSF) to enhance above model to

obtain a sophisticated segmentation method. In former manuscript, we cannot give the clear and detailed description about proposed model, so in revised paper, we first give the corrected cite of previous work, and then provide clear explanation about the goal of our work in line 7-27 of page 2, the enhanced theory about proposed SCM-motivated enhanced CV model can be seen in subsection “C. Mass Candidate Localization ” and subsection “D. Segmentation Methods”, and thank you very much for your reminder, these changes let the theory of our method more clear and readable. Comment 2: You may avoid using abbreviations in the title of the article Thank you for your comment, using abbreviations to make the title of our article more short and straightforward, and we provide the full name of all the abbreviations at the first time, and thank you for your reminder again. Comment 3: While using the abbreviations at the first time, write their expansions, for example, SCM, CV, and PCNN. Thank you for your comment, we added all the full name of SCM, CV and PCNN model at the first time as you suggested. Comment 4: In table 3, first row, the sensitivity is reduced for the proposed algorithm compared to the method 2. For all the other examples it has increased. Is there any s Use abbreviations to make the title of the article more straightforward pecific reason for that?

TP---True Positives

FP---False Positives

FN---False Negatives

Contour of ground truth

Contour of the method

fig.3 definition of TP, FP and FN.

Thank you for your comment, according to the definition of Sensitivity:

SEN 

TP TP  FN

Here, we define the areas of Ground truth as GT, from the fig.3, we can achieve that GT=TP+FN, that’s, FN=GT-TP, then SEN 

TP TP  FN



TP TP  (GT  TP )



TP

. In normal case, GT is a constant, so the Sensitivity is

GT

proportional to TP. For image a, we can see from (a-3_enlarged) for Method 2 and (a-4_enlarged) for proposed method, the TP areas of Method 2 is great larger than proposed method, so the sensitivity of

Method 2 is higher than proposed method, but the Specificity of our method is almost 1.93% high than Method 2, that’s to say, neither sensitivity nor specificity can be very objective in evaluating algorithmic performance, so in final step, AUC is introduced to measure the method combining sensitivity and specificity. As a matter of fact, the title of paper should avoid abbreviations, we apply the abbreviations such as SCM and CV to make title more short and straightforward, and the “SCM-motivated Enhanced CV Model for Mass segmentation from Coarse-to-Fine in Digital Mammography” might be straightforward and clear enough to explain the aim of our work, and thank you a million. Comment 5: The sentence immediately after table 7 has an improper formation. Thank you for your comment, and we have added the proper formation, thank you for your reminder. Comment 6: What kind of subjective measure have you used to identify the affected area in mammogram? Why have you used only experts to find the area and not naive viewers? Thank you for your comment, massive experiments are conducted on two common and public databases including DDSM and MIAS, so we just can use the ground truth provided by experts to measure the segmentation capability of proposed method. In future work, we will make the cooperation with breast cancer hospital of Gansu and can obtain more clinical mammogram to verify our model for mass segmentation, and what you said can be considered in our future work, thank you again.

In conclusion, thank you for your critical comments and we totally agree with your suggestions, these comments are of great help to improve the quality of our manuscript. We appreciate for yours’ warm work earnestly, and hope that the correction will meet with approval. Once again, thank you very much for the valuable comments and suggestions. We hope everything goes well and look forward to hearing from you. Sincerely yours Yide Ma Address: No. 222, South Tianshui Road, Lanzhou, Gansu Province, 730000, P.R. China. Affiliation: School of Information Science and Engineering, Lanzhou University. E-mail: [email protected], [email protected].

Manuscript

Click here to download Manuscript Revised paper.docx

Click here to view linked References 1 2 SCM-motivated Enhanced CV Model for Mass segmentation from Coarse-to-Fine in 3 4 Digital Mammography 5 6 7 Ya’nan Guo, Xiaoli Gao, Zhen Yang, Jing lian, Shiqiang Du, Huaiqing Zhang and Yide Ma1 8 9 10 Lanzhou University, School of Information Science and Engineering, Lanzhou, Gansu, China 11 12 Abstract 13 14 A novel approach for mass segmentation from coarse-to-fine in digital mammography, termed as SCM-motivated enhanced 15 CV algorithm, is presented in this paper. As well known, it is difficult to robustly achieve mammogram mass segmentation due to 16 17 low contrast between normal and lesion tissues, as well as high density tissue interference in mammograms. Therefore, Spiking 18 Cortical Model with biology background is introduced to achieve mammary-specific and mass edge detection, and this mass 19 20 candidate is regarded as the initial contour of improved CV model followed by, effectively overcoming the drawback that CV 21 method is sensitive to the initial contour; especially, the enhanced CV model innovatively combines the techniques of physical 22 23 imaging principle, and local region-scalable force, harvesting the coarse-to-fine mass boundary accurately. The proposed method 24 is tested totally on 400 mammograms from two well-known digitized datasets (digital database for screening mammography and 25 26 mammography image analysis society database), achieving the average detection rate of 93.25%. By comparing with the region27 based model with bias field (Method 1) and typical CV model (Method 2), we can reach the conclusion that proposed method is 28 29 outperform other methods, yielding the average sensitivity of 95.83%, specificity of 99.13%, dice similarity co-efficient of 92.21% 30 and AUC of 98.02%. In addition, this method is verified on the mammograms from Gansu Provincial Cancer Hospital, the detection 31 32 results reveal that our method can accurately detect the abnormal in clinical application. 33 Index Terms—Mammography, Mass Segmentation, Spiking Cortical Model (SCM), Enhanced CV model, Local Region34 35 Scalable Force (LRSF). 36 37 38 I. INTRODUCTION 39 Recent statistics show that breast cancer is one of the common cancer with a high mortality rate among women [1,2]. Complete 40 41 curing of breast disease is possible once it is detected in early stage, specifically, early detection and diagnosis can achieve the 42 survival rate of 95%. Mammography is currently regarded as the best choice for early detection of breast cancer in women, which 43 44 can depict most of the significant changes of breast disease [3] and is a specific type of imaging that employs a low-dose X-ray 45 system and high-contrast, high-resolution film for breast examination. The main problem to analysis mammogram is that low 46 47 contrast between normal and lesions tissues, blurry and irregular lesions shape, high density tissue interference and much noise in 48 49 such images, which form a challenge for radiologists to make diagnosis decision. Masses are one of the most common signs in 50 mammograms [4], accurate mass segmentation is the fundamental and significant tasks in breast cancer diagnosis. 51 52 As is known to all, accuracy segmentation of mass plays an important role for mass classification, since most valuable masses 53 properties are related to its morphology, which is used to judge whether malignant or not. American College of Radiology 54 55 (ACR) also defines a standard for mammography reporting named BI-RADS (Breast Imaging Reporting and Data System) [4] to 56 57 58 59 60 1 E-mail address: [email protected], [email protected]. 61 Postal address: No.222. Tianshui Road (South). Lanzhou. Gansu Province. China. 730000. 62 Tel: (+86)-139-9311-2999; Fax: 09318912786 63 1 64 65

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

evaluate the lesions grade. In the last years, an amount of mass segmentation algorithms were proposed to assist radiologists for early diagnosis and treatment [5], which mainly included: the segmentation based on regions [6-8], the segmentation based on contours [9-11] and the segmentation with thresholding [12-14] et.al. In recent years, many of work on active contour implemented via level set methods have been proposed to address a wide range of image segmentation problems in image processing and computer vision [15,16], among them, Chan and Vese (CV) proposed an active contour model by a variational level set formulation [17], this typical CV model performs more excellent characteristics in boundaries segmentation with the capability of convergence and superior noise robustness. Despite many good numerical results obtained with above CV model and strong theoretical properties, it still exists some intrinsic limitations. Its flaws mainly reflect in two aspects: first, typical CV model is proposed based on the assumption that the intensities in each region is very close to a certain constant, so it might lead to a poor segmentation result when extending to segment image with inhomogeneous object; second, traditional CV model is highly sensitive to the initial condition, that’s to say, different initial contour might result in different segmented results for same image, so the choice of the initial contour is vital to promising results. In response to the problem of reinitialization of CV model, many efficient implementation schemes have been proposed in [18-22]; as to efficiently perform the segmentation of images with intensity inhomogeneity, a new class of models has been proposed, which can be found in literatures [23-26]; simultaneously, some powerful methods were recommended to solve these two limitations, we can find them in papers [27-29], although above approaches achieved competitive detection results, there is still room for improvements with the development of technology. In this work, what we most concerned about is how to handle the dilemmas of CV model when extending to segment the mass with low-contrast and blurry boundaries in mammograms. Based on the aim of mass detection is to segment the mass region from an image, SCM-motivated two-phase enhanced CV model is presented, in which Spiking Cortical Model (SCM) model [30] that is in line with human visual characteristics is introduced to achieve the initial contour, and the principle of physical imaging is used to simulate bias field [31] for intensity correction, for refinement of mass segmentation, local region-scalable external force (LRSF) [32] is recommended to improve the internal energy evolution mechanism in weak edge of object. The proposed model is tested on 400 mammograms from digital database for screening mammography (DDSM) and mammography image analysis society database (MIAS), and the experimental results prove its effectiveness by making comparisons with region-based model with bias field [31] and typical CV model [33]. As well, subjective and objective evaluation results exhibit its superior segmentation performance for mammograms, Note that this paper is an extended version of our preliminary work presented in our conference paper [29]. This paper is organized as follows: In Section II, firstly, we introduce the selected databases, and then we present the method of pre-processing and mass coarse location. After that, proposed methodology is briefly described. Section III illustrates massive experiments to verify the proposed method. Besides, the comparisons and the discussion can be found in this Section. Section IV gives the conclusions of this paper. II. METHODOLOGY In this work, mass segmentation includes four steps: 1) Input the mammogram; 2) Mammogram pre-processing, remove the label and pectoral muscle, and then enhance image; 3) Mass rough location, employ SCM model to obtain mass rough contour; 4) Mass fine segmentation, SCM-motivated enhanced CV model is proposed. A. Image Database In this work, a set of images selected from Mammography Image Analysis Society (MIAS) database [34] and the digital database for screening mammography (DDSM) [35] are used for algorithm verification. The DDSM consists of 2620 cases, 2

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

available in 43 volumes. A case is a collection of mammograms and information corresponding to one mammography exam of one patient. The DDSM database provides the chain codes of the suspicious regions and metadata of each abnormality using the BI-RADS lexicon, and it contains the ground truth information of the abnormalities locations. The MIAS database provides 322 mammograms. Every mammogram is 1024 × 1024 pixels, it is provided the corresponding information of lesion area such as type, location, severity and character of background tissue. The mass database consists of 64 images. The tumors have various shape and size, the ground truth is the central coordinate and radius drew by experts[36]. B. Pre-processing The pre-processing operation is required due to its low contrast. The aim of mammogram pre-processing is to remove the labels, pectoral muscle and noise, and then enhance the contrast of lesion area. 1) Label and pectoral muscle removal The mammogram usually includes labels and pectoral muscle, in order to reduce its interference, the labels and muscle should be removed firstly. We apply the methods of selecting the largest connected region and seed region growing algorithm [34] to remove the labels and muscle respectively. 2) Image Enhancement After that, the obtained image is enhanced to improve the contrast by the nonlinear unsharp masking algorithm (NLUM) [37]. Through this operation, the local contrast of specific regions and fine details are effectively improved, more importantly, this preprocessing operation makes the brightness variation more obvious between the mass and surrounding tissue, facilitating the substantial mass segmentation. C. Mass Candidate Localization As known, the typical CV model is quite sensitive to the initial conditions, so we need to obtain the approximate mass location as the initial contour for more accurate segmentation. SCM [30] is another improved version of Pulse Coupled Neural Network (PCNN) model [38,39], which has better biological background by combining the advantages of both PCNN and intersecting Cortical Model (ICM) [40]; it is very suitable for image processing because it has a high sensitivity for low intensities of stimulus but low sensitivity for high intensities, more importantly, SCM has lower computational complexity compared to PCNN, so we apply SCM to achieve mass location, the SCM model can be deprived and expressed as: U ij ( n)  fU ij ( n  1)  Sij  Wijkl Ykl (n  1)  Sij kl

(1)

Eij ( n )  gEij ( n  1)  hYij ( n  1)

(2)

1,1/(1exp(  (Uij ( n )Eij ( n ))))0.5 Y ij ( n )   0,otherwise 

(3)

h g

Yij Eij

W Sij Sij

Σ

U ij

O

I

Yij

f

Fig.1. SCM model: in this model, assuming the feedback input equals the external stimulus, the linking input equals the convolution of the pulse modulation mechanism; internal activities are modulated nonlinear not only by the input field and the linking field, but also by the decay of memory neuron state; The linking field is determined by the synaptic strength coefficient, and the threshold is attenuated by exponential rule, the similar neurons can release pulses simultaneously. In the same way, for image segmentation, the same or similar grayscale pixel points will ignite simultaneously so that the target segmentation can be achieved.

3

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

Where

U ij ( n) is

internal activity,

applied to linking field, f and

Sij

is a stimulus,

Yij (n)

is output,

Eij ( n)

is dynamic threshold, Wijkl is synaptic weight matrix

g denote decay constants, h stands for threshold magnitude coefficient,  denotes a parameter of

sigmoid function. The nonlinearity of sigmoid function can be used to generate pulse. Obviously, SCM has a more compact structure than PCNN models, which can be seen in Fig.1, so it decreases the computation complexity. In this work, the initial mass contour is obtained by SCM model, through massive tests, the parameters are set as follows: U ij (0)

W 

initialization is set as 0, and

 0.1091  0.1409   0.1091

0.1409

0.1091

0

0.1409

0.1409

0.1091

.   

Eij (0)

initialization is set as 1, f=0.2, g=0.9, h=20,

 =1. Synaptic weight matrix is set as

The image pre-processing process and mass lactation results are given in Fig.2.

(a)

(b)

(c)

(d)

(e)

Fig.2. Pre-processing and localization results:(a) Original image (b) label removed image (c) muscle removed image (d) enhanced image (e) mass localization by SCM model.

D. Segmentation Methods SCM model is employed to obtain the mass contour, but there exists certain gap with the mass actual boundary. So we utilize original CV model to perform the accurate segmentation further, but the experimental results denote that typical CV model may not work well on mammograms because of ambiguous masses margins, that’s, the normal regions around the masses always present the similar characteristics to masses. Based on above problem, a novel SCM-motivated enhanced CV model is proposed via the theory of the physical imaging principle and local region-scalable force to segment the mass in mammogram. 1) Typical CV Model In our work, we just consider the case with one mass in mammogram, that’s, we divide the image into two regions including mass and background, here, two-phase level set formulation was introduced to simply the Mumford-Shah functional by Chan and Vese [33], the energy function of partitioning an image I ( x, y )   into two regions (mass and background) were expressed as:

 

 

 

F CV  , c1, c2    I  x   c1 H   x  dx  I  x   c2 1 H   x   dx V  H   x  dx 2









(4)

Where the first two terms denote the date fitting terms. c1 , c 2 are the averages of inside and outside region of contour in image I respectively. The last term denotes the regulation term, regulating the length of level set curve. H () is the Heaviside function.  ( x ) is the level set function, V denotes a parameter that represents the length of the contour. At this point, mass segmentation is translated into the math problem of solving  ( x ) , c1 and c 2 that minimizes above energy function F CV . Although above typical CV model has been widely applied to solve many practical problems as long as the image intensity are very close to piecewise constant, it faces great challenge when extending to segment the image with inhomogeneous object, just because which is beyond its original assumption [26]. Base on inhomogeneous intensities in image, Li et al. introduced the principle of physical imaging simulate bias field estimation [31], which can give certain correction of image intensities and improve the accuracy for image segmentation. In following part, we will give the theory of this model.

4

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

2) Physical Imaging Principle According to the principle of physical imaging, the observed image I in real-world can be modeled as:

I  bJ  n

(5)

Where J denotes the true image, b is the component that accounts for the intensity inhomogeneity, here, it is referred as a bias field, and n denotes additive noise. According to the features of real image and bias field b , we make the following assumptions: (A1) The bias field b is slowly varying, which implies that each pixel can be approximated by a constant in the neighborhood domain of the image. (A2)The true image J can be divided into N disjoint regions 1, C1

CN

N

, which can be approximated by N different constants

.

3) Energy Formulation Based on Bias Field At each point y in image I , we define a circular neighborhood

with radius  . According to the image model

O y  {x : x  y  }

in (5) and above two assumptions, when bias field b is varying slowly, the values of b( x ) are very close to b (y) in this circular neighborhood O y  { x :

x  y  } ,

i.e.

b( x)  b(y)

for x  O y

Thus, b( x ) J( x ) are close to a constant b (y) C i in each sub-region O

b( x) J( x)  b(y) Ci

y

(6) , that’s:

 i

for x  O y

i

(7)

According to image model in (5), we can achieve:

I ( x)  b(y) Ci +n( x)

for x  O y

i

(8)

Here, n( x ) denotes additive zero-mean Gaussian noise, so the intensities in above set are described as:

i I y  { I ( x ) : x  O y  i } Obviously, each subregion image can be divided into N clusters of

1 Iy,

N , Iy

(9) , and



b y Ci , i  1,

, N can be regarded as its

cluster centers, this local intensity clustering property is somewhat similar to clustering algorithm, so here K-means clustering algorithm is applied to achieve the iterative process to minimize the clustering criterion in the neighborhood. Through above, the energy function of improved CV model can be rewritten in a continuous form as follows:





F CV     K  y  x  I  x   b  y  Ci dyM i   x   dx  V  H   x  dx N

2



i 1

For convenience, let C denotes the constants with a vector C  (C1 ,

,C

N

),

(10)

the energy function can be expressed by the

variables of level set function  ( x ) , the vector C , and the bias field b . For mass segmentation, two-phase CV model is utilized, that’s, N  2 , its membership functions are

   H   and M 2    1  H   . K ( y  x) denotes as a nonnegative window

M1 

function, which is also a kernel function, if K ( y  x )  0 , then x  Oy , more detailed information can be found in [31]. Then formulation in (10) can be rewritten as:

5

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

 K ( y  x ) I  x   b ( y ) C1  2  K （y  x）I  x   b ( y ) C 2 1  H   x  F

CV

 , C, b  





2





H   x  dx

 dx  V

(11)





 H   x  dx 

4) Enhanced CV Model Although the mass contour can be obtained when above SCM-motivated CV method with bias field is applied, there exists obvious distortions in obtained results, that’s, the obtained contour cannot completely converges to the target boundary where the contrast is relatively low and boundary is blurry. Thus, we introduce a local region-scalable external force [32] to enhance above model. Through this, the mass contour can converge well even in weak boundary.

Fig.3 Scalable local region on expandable kernel

Specific implementation of LRSF is to add a window function in the contour line. As shown in Fig.3, the centers of these window functions are located on the contour and evenly distributed along the contour, the size of window function is scalable. The contour divides these windows into two parts including inside and outside of contour, and we respectively conduct the greyscale statistical calculation for the inside and outside region of the contours. If statistical gray level difference of the inside and outside the contour is very small or approximately equal zero in a certain window function, we believe that this window is in a homogeneous region, at this time, the contour is not the actual boundary of the object, and the pressure generated by these difference is approximately zero, so the contour line does not converge to any edge, adjusting the size of the window function to make sure both target and background region are included, and then do the statistical gray as before, in this process, we let window cover these two areas by controlling the size of l , when the gray difference between these two areas is relatively large, the evolution direction of the contour line can be judged, and the real boundary can be found. In mammogram, the breast mass is the brightest region in the image, which means that the target area has a larger gray area, in contrast, background has smaller ones, and the actual edge of the mass can be achieved only if the contour line evolves from higher intensities area to lower one, That is, the contour line moves to the direction with higher gray value. In the end, as long as gray level statistics difference of inside and outside region is large enough, even a very small window function is selected, our method still can converges to the true boundary breast lumps. Here, Gaussian kernel function is used to calculate regional statistical results, these two local regional statistics are the weighted average intensity in and outside the contour respectively, defined as u and v , which can be written as: u  y 

 K  x  y H   x   I  x  dx  K  x  y H   x   dx 6

(12)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

v y 

 K  x  y  1 H   x  I  x  dx  K  x  y  1 H   x  dx

(13)

Then, the energy function of the proposed enhanced CV model can be rewritten as: F CV  , b  

 K ( y  x) I  x   b( y )u ( y )



 K （y  x）I  x   b( y )v ( y ) 

2

2

H   x   dx

(14)

1  H   x    dx  V  H   x   dx 

By minimizing above energy function [17], we can get the fine and accurate boundary of the mass in mammogram. III. EXPERIMENTS AND QUANTITATIVE ANALYSIS In this section, we run a set of experimentations on DDSM and MIAS database, and compare the performance of proposed model with region-based model with bias field [31] (Method 1) and typical CV model [33] (Method 2), in total 400 mammograms in which 64 mammograms are from MIAS database and 336 mammograms are from DDSM database, and each case contains one abnormality. The visual evaluation are shown in Fig.4 and Fig.5 respectively. A. Subjective Evaluation In 336 DDSM images, 315 images can obtain the accurate mass contour, in which 300 images can fully converge to mass boundary. For DDSM images, the position information of the masses have been given by chain code date as ground truth. The thumbnail images for visual browsing of each cases are shown in Fig.4-(a)-(d). Fig.4-(a-5)-(d-5) shows the ground truth of mass that were marked as the red curves by experts. The images converted into LJPEG format are shown in Fig.4-(a-1)-(d-1). The segmentation results obtained by proposed method are shown Fig.4-(a-4)-(d-4), and the enlarged results are shown in Fig.4-(a4_enlarged)-(d-4_enlarged). As can be seen, the contours obtained by proposed method can accurately converge to mass boundary in all cases. From the enlarged images, we can clearly see the contour are very close to the ground truth shown in Fig.4-(a-5)-(d5). The segmentation results of Method 1 and Method 2 are shown in (a-2)-(d-2) and (a-3)-(d-3) of Fig.4, Fig.4-(a-2_enlarged)-(d2_enlarged) and Fig.4-(a-3_enlarged)-(d-3_enlarged) denote the enlarged results respectively. From the enlarged results, we can notice these two methods cannot completely converge to the target boundary even the initial contour is very close to the true boundary, they are all easy to be affected by blood vessels and other texture region around the mass; it is obvious that Method 1 has the better segmentation capability than Method 2, that’s to say, the region-based model with bias field has relatively better convergence ability than typical CV method; Fortunately, proposed SCM-motivated enhanced CV model not only has improved the diploma the sensitivity of initial contour of CV, but also stronger capability of convergence, that is, our approach is stable enough and has strong robustness, performing better than other two methods for DDSM database.

(a)

(b)

(c)

7

(d)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

(a-1)

(a-2)

(b-1)

(b-2)

(a-2_enlarged)

(b-2_enlarged)

(a-3)

(b-3)

(a-3_enlarged)

(a-4)

( c-1)

(c-2)

(d-1)

(d-2)

(c-2_enlarged) (d-2_enlarged)

(c-3)

(b-3_enlarged)

(b-4)

(c-3_enlarged) (d-3_enlarged)

(c-4)

8

(d-3)

(d-4)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

(a-4_enlarged)

(a-5)

(b-4_enlarged)

(c-4_enlarged) (d-4_enlarged)

(b-5)

(c-5)

(d-5)

Fig.4 Experimental results of mammograms from DDSM database: (a)-(d) The”thumbnails”; (a-1)-(d-1) Full raw images; (a-2)-(d-2)Results by Method 1;(a-3)(d-3)Results by Method 2; (a-4)-(d-4)Results by proposed method; (a-5)-(d-5)The ground truth.

In 64 MIAS images, 58 images can obtain the accurate mass location by comparing with the given standard. Here, the segmentation results of four images including mdb010, mdb015, mdb028 and mdb184 are displayed in Fig.5, respectively. In Fig.5-(e-4)-(h-4), blue contour denotes accurate ground truth segmented manually and the red circles contour denotes the coarse ground truth provided by expertise, here, we aim at exploring the more infallible mass contour, so ground truth segmented manually are utilized to measure the segmentation capability of different approaches. For MIAS images, we first apply the proposed preprocessing method to manipulate mammogram, and then, novel SCM-motivated enhanced CV model is used to achieve the precise mass contour. The results of proposed method are shown in Fig.5-(e-3)-(h-3), and the enlarged results are shown as Fig.5-(e3_enlarged)-(h-3_enlarged); The Fig.5-(e-1)-(h-1) and Fig.5- (e-2)-(h-2) display the comparison results of Method 1 and Method 2. The Fig.5-(e-1_enlarged)-(h-1_enlarged) and Fig.5-(e-2_enlarged)-(h-2_enlarged) are the enlarged results of these two test methods respectively. Compared with these two models, proposed method can completely remove the labels and interference, and can achieve more robust and accurate results. Even in blurry region in mdb028 and mdb184, our method still can approach to real edge. The mass in mdb028 and mdb184 are malignant, so the margin of mass is rough; the mass in mdb010, mdb015 are benign, so the mass edge is relatively smooth. Our results can objectively reflect the pathology characteristics of actual masses to some extent that the malignant masses are always with burrs. This performance somewhat benefits to the early diagnosis of breast cancer and also indicates the superiority of our method. Method 1 holds the better visual results than Method 2.

(e)

(f)

(g)

9

(h)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

(e-1)

(f-1)

(e-1_enlarged)

(f-1_enlarged)

(e-2)

(g-1_enlarged)

(f-2)

(e-2_enlarged)

(f-2_enlarged)

(e-3)

(f-3)

(e-3_enlarged)

(g-1)

(g-2)

(g-2_enlarged)

(g-3)

(f-3_enlarged)

(g-3_enlarged)

10

( h-1)

( h-1_enlarged)

(h-2)

( h-2_enlarged)

(h-3)

( h-3_enlarged)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

(e-4)

(f-4)

(g-4)

(h-4)

Fig.5 Experimental results of mammograms from MIAS database: (e)-(h) Full raw images; (e-1)-(h-1) Results by Method 1; (e-2)-(h-2)Results by Method 2;(e3)-(h-3) Results by proposed method; (e-4)-(h-4) The ground truth.

B. Quantitative Analysis Choosing the right evaluation metric is crucial for performance evaluation, we need right metric to compare their segmentation ability with others [41]. If it need the ground truth, we call it as supervised evaluation metric; otherwise we call it as unsupervised evaluation metric. In our paper,we choose the supervised evaluation metrics [42] as follows: 1) Detection Rate The detection rate is defined as the ratio of actual detected mass number and total mass number. The detection rates of each databases are shown in Table1, from Table 1, we can notice 373 images are detected successfully, therefore, the average detection rate is 93.25%, the detection rate for DDSM database is up to 93.75%, but the detection rate in MIAS is just 90.63%, this is just because the lesions of dense breast in MIAS database are always embedded in the gland, and the boundary of the masses are blurry, so the mass contour are hard to obtain. Therefore, the detection rate is relatively lower. Table 1 Detection rate Database

Tested images

DDSM

336

MIAS

64

Total

400

Detected images

Non-detected images

Detection rate(%)

315

21

93.75

58

6

90.63

373

27

93.25

2) Area Overlap Metric In order to quantitatively evaluate the mass segmentation results, we introduce the area overlap metric :

O Where

L

L T L T

(12)

is the mass area segmented by a particular method, and

intersection area of

L

and T ;

L T

T

is the mass area of ground truth. L  T denotes the

is union set area of these two regions. The value of

means these two regions have no overlap;

O =1

O

is bound between zero and one.

means these two regions exactly overlap. The higher value of

O

O =0

means better

detection result. Mean and variance of the area overlap metric denote the stability of the algorithm, higher mean value and lower variance means the better performance of the algorithm. We list the average area overlap metric and variance of these three segmentation methods respectively in Table 2. Table 2 The area overlap ratios Method

Mean (%)

Variance (%)

Method 1

84.154

4.7958

Method 2

76.343

7.5498

Proposed method

89.192

3.4641

As is shown in Table 2, the average area overlap ratio of proposed method is obviously higher than the other two methods, as well, the variance is the lowest among them. In conclusion, our segmented results are much more close to the ground truth; the segmentation ability of Method 1 ranks second, Method 2 just employs the original CV model, so its performance is not better 11

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

enough. 3) Comparison by Similarity Measure a) Sensitivity and Specificity The sensitivity and specificity [43] refer to the ability of the method to correctly identify those masses in mammogram. The sensitivity and specificity are written as follows: SEN 

TP TP  FN

(13)

SPE 

TN TN  FP

(14)

Definitions of TP, TN, FP and FN are as follows: TP: the region is abnormal and the tests segment it as abnormal region. FP: the region is normal and the tests segment it as abnormal region. TN:the region is normal and the tests segment it as normal region. FN: the region is abnormal and the tests segmente it as normal region. The sensitivity and specificity values of the eight images of (a)-(d) and (e)-(f) are shown as Table 3 and Table 4 Table 3 Sensitivity Method 1 0.9159±0.0194 0.8574±0.0387 0.9304±0.0251 0.8973±0.0106 0.8695±0.0326 0.7709±0.0164 0.7341±0.0347 0.9575±0.0369

a b c d e f g h

Method 2 0.9849±0.0056 0.9550±0.0290 0.9217±0.0278 0.9956±0.0022 0.5793±0.0347 0.5709±0.0193 0.6766±0.0315 0.8841±0.0442

Proposed method 0.9667±0.0022 0.9553±0.0288 0.9601±0.0133 0.9997±0.0004 0.9838±0.0080 0.9371±0.0152 0.8382±0.0382 0.9952±0.0090

Table 4 Specificity Method 1 0.9560±0.0043 0.9611±0.0025 0.9522±0.0075 0.9460±0.0011 0.9696±0.0044 0.9853±0.0044 0.9991±0.0029 0.7151±0.0414

a b c d e f g h

Method 2 0.9774±0.0040 0.9614±0.0025 0.9668±0.0228 0.9656±0.0011 0.9780±0.0174 0.9889±0.0062 0.9902±0.0110 0.9542±0.0199

Proposed method 0.9967±0.0023 0.9949±0.0012 0.9817±0.0041 0.9919±0.0007 0.9996±0.0006 0.9977±0.0046 0.9998±0.0010 0.9684±0.0114

b) Dice Similarity Co-efficient The dice similarity co-efficient (DSC) is another evaluation metric used to compare the similarity between the segmented contour and the ground truth. It is defined as: DSC 

2* L T 2 *TP  LT 2 * TP  FN  FP

(15)

The DSC is valued between zero and one. From the equation, we know that the bigger value means better segmentation performance. The results of the Dice Similarity Co-efficient are shown as Table 5:

12

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

Table 5 The Dice Similarity Co-efficient Method 1 0.8313±0.0207 0.8671±0.0046 0.8729±0.0160 0.7592±0.0065 0.8223±0.0286 0.8479±0.0074 0.8157±0.0218 0.8803±0.0221

a b c d e f g h

Method 2 0.8985±0.0176 0.8677±0.0051 0.9154±0.0063 0.8261±0.0127 0.7305±0.0257 0.7423±0.0303 0.8059±0.0316 0.9217±0.0215

Proposed method 0.9410±0.0047 0.9069±0.0217 0.9228±0.0157 0.9008±0.0076 0.9165±0.0056 0.9432±0.0051 0.8888±0.0160 0.9569±0.0139

Table 6 The average sensitivity, specificity and Dice Similarity Co-efficient Evaluation metric

Method 1

Method 2

Proposed method

Sensitivity

0.8666

0.8210

0.9583

specificity

0.9356

0.9728

0.9913

Dice Similarity Co-efficient

0.8385

0.8371

0.9221

Table 3, Table 4 and Table 5 display the evaluation of sensitivity, specificity and the DSC for eight images of (a)-(d) and (e)(f), in order to give the comprehensive and objective evaluation, we average all these three indicators, and Table 6 shows their average vaules. We can notice proposed method holds the highest average vaules of 95.83% sensitivity, 99.13% specificity and 92.21% DSC, which infers its better mass segmentation capcibity, moreover, our method can be more close to the ground truth to the maximum extent in all cases. Method 1 have the relatively higher DSC value than Method 2. 4) Performance Analysis by ROC Curves Ultimately, Receiver Operating Characteristics (ROC) curve is introduced to measure the predictive accuracy of the proposed model comprehensively. The area under the curve (AUC) is regarded as a descriptor of test results. ROC curve for three methods are shown in Fig.6. The result of AUC is shown in Table 7.

Fig.6 ROC curves Table 7 The AUC of the methods Method

Method 1

Method 2

Proposed method

AUC

0.9246

0.9184

0.9802

From Table 7, we discovery the AUC of our model is up to 0.9802, which is 5.56% higher than Method 2 and 6.18% higher than Method 1, that is to say, the proposed method is outperform the others. Method 1 holds the relatively high AUC of 0.9349, so it ranks second; Method 1 ranks third with the AUC of 0.9045. Based on Table 3, 4, 5, 6, 7 and the Fig.6, we can make the conclusion that the performance of proposed method outperforms other methods, yielding the average sensitivity of 95.83%, specificity of 99.13%, DSC of 92.21% and AUC of 98.02%. 5) Clinical Verification To verify the validity of the proposed method in clinical application, we implement it to the mammograms from Gansu 13

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

Provincial Cancer Hospital. The experimental results are shown in Fig.7, we discover that our algorithm can detect the abnormal accurately.

(a) Original image; (b) Region of Interest; (c) Detected abnormal Region by SCM-motivated improved CV model.

Fig.7. Results of Clinical validation

IV. CONCLUSIONS In this paper, a novel SCM-motivated enhanced two-phase CV model is proposed for mass fine segmentation. The proposed model regards the segmentation result of SCM model as the initial contour of improved CV, which greatly improves the phenomenon that traditional CV model is sensitive to the initial contour; moreover, we introduce the strategies of the physical imaging principle and local region-scalable external force to enhance the performance of CV model, the proposed model not only inherits the advantage of original CV model, but also improve its stability and robustness. The quantitative and qualitative evaluation demonstrate the effectiveness of the method, and our method can converge to the mass contour even in blurry regions. By comparing with other methods, we have the conclusion that our approach outperforms other methods, yielding the average sensitivity of 95.83%, specificity of 99.13%, DSC of 92.21% and ROC of 98.02%. Propose model can achieve better mass segmentation results, and we firmly believe that it can be recommended to clinical application for various lumps. In our future work, we also try to introduce the theory of our method into the widespread application for event analysis [44], feature analysis [45], multiclass discriminant analysis [46] and visual tracking [47-50], more open-mind and fresh ideas also be welcomed to enhance this model for other different occasions. ACKNOWLEDGMENTS Authors would like to thank the retrieval of all the public database for the experiments of this paper. This study was funded by the National Natural Science Foundation of China (nos.61175012 and 61201421) and Natural Science Foundation of Gansu Province (nos. 145RJZA181 and 1208RJZA265). CONFLICT OF INTEREST The authors declare that they have no conflict of interest. REFERENCE 1. Jemal A, Bray F, Center MM, Ferlay JJ, Ward E, Forman D (2011) Global cancer statistics. CA. Cancer J Clin. Ca A Cancer Journal for Clinicians 61 (2):69-90 2. Xie W, Li Y, Ma Y (2016) PCNN-based level set method of automatic mammographic image segmentation. Optik - International Journal for Light and Electron Optics 127 (4):1644-1650 3. Timp S, Varela C, Karssemeijer N (2007) Temporal change analysis for characterization of mass lesions in mammography. IEEE Transactions on Medical Imaging 26 (7):945-953 4. Edwards SD, Lipson JA, Ikeda DM, Lee JM (2013) Updates and revisions to the BI-RADS magnetic resonance imaging lexicon. Magnetic Resonance Imaging Clinics of North America 21 (3):483-493 5. Salmeri M, Mencattini A, Rabottino G, Accattatis A, Lojacono R Assisted Breast Cancer Diagnosis Environment: A Tool for DICOM mammographic images analysis. In: Medical Measurements and Applications, 2009. MeMeA 2009. IEEE International Workshop on, 2009. pp 160-165 6. Elter M, Held C, Wittenberg T (2010) Contour tracing for segmentation of mammographic masses. Physics in Medicine & Biology 55 (55):5299-5315

14

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

7. Kim DH, Choi JY, Ro YM (2014) Region based stellate features combined with variable selection using AdaBoost learning in mammographic computer-aided detection. Computers in Biology & Medicine 63 (C):238 8. Eches O, Benediktsson JA, Dobigeon N, Tourneret JY Adaptive Markov random fields for joint unmixing and segmentation of hyperspectral images. In: Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, 2010. pp 1-4 9. Berber T, Alpkocak A, Balci P, Dicle O (2013) Breast mass contour segmentation algorithm in digital mammograms. Computer Methods & Programs in Biomedicine 110 (2):150-159 10. Liu J, Liu X, Chen J, Tang J Mass Segmentation in Mammograms Based on Improved Level Set and Watershed Algorithm. In: Advanced Intelligent Computing Theories and Applications. with Aspects of Artificial Intelligence - International Conference, Icic 2011, Zhengzhou, China, August 11-14, 2011, Revised Selected Papers, 2011. pp 502-508 11. Lu X, Dong M, Ma Y (2015) Automatic Mass Segmentation Method in Mammograms Based on Improved VFC Snake Model, Elsevier Inc, pp. 201-217. 12. Wei K, Guangzhi W, Hui D Segmentation of the breast region in mammograms using watershed transformation. In: IEEE Engineering in Medicine and Biology Conference, 2006. pp 6500-6503 13. Hsu WY (2012) Improved watershed transform for tumor segmentation: Application to mammogram image compression. Expert Systems with Applications 39 (4):3950-3955 14. Liu CC, Tsai CY, Liu J, Yu CY, Yu SS (2012) A pectoral muscle segmentation algorithm for digital mammograms using Otsu thresholding and multiple regression analysis. Computers & Mathematics with Applications 64 (64):1100–1107 15. Dera D, Bouaynaya N, Fathallah-Shaykh HM (2016) Automated Robust Image Segmentation: Level Set Method Using Nonnegative Matrix Factorization with Application to Brain MRI. Bulletin of Mathematical Biology 78 (7):1-27 16. Liu C, Liu W, Xing W (2017) An improved edge-based level set method combining local regional fitting information for noisy image segmentation. Signal Processing 130:12-21 17. Chan TF, Vese LA (2001) Active contours without edges. IEEE Transactions on Image Processing A Publication of the IEEE Signal Processing Society 10 (2):266-277 18. Solem JE, Overgaard NC, Heyden A Initialization Techniques for Segmentation with the Chan-Vese Model. In: International Conference on Pattern Recognition, 2006. pp 171-174 19. Xia R, Liu W, Zhao J, Li L An Optimal Initialization Technique for Improving the Segmentation Performance of Chan-Vese Model. In: IEEE International Conference on Automation and Logistics, 2007. pp 411-415 20. Li C, Xu C, Gui C, Fox MD (2010) Distance regularized level set evolution and its application to image segmentation. IEEE Transactions on Image Processing A Publication of the IEEE Signal Processing Society 19 (12):3243 21. Niu S, De SL, Chen Q, Leng T, Rubin DL (2016) Automated geographic atrophy segmentation for SD-OCT images using region-based C-V model via local similarity factor. Biomedical Optics Express 7 (2):581 22. Wu Y, Hou W, Wu S (2011) Brain MRI segmentation using KFCM and Chan-Vese model. Transactions of Tianjin University 17 (3):215-219 23. Piovano J, Rousson M, Papadopoulo T, odore Efficient segmentation of piecewise smooth images. In: International Conference on Scale Space and Variational Methods in Computer Vision, 2007. pp 709-720 24. Lankton S, Tannenbaum A (2008) Localizing region-based active contours. IEEE Transactions on Image Processing A Publication of the IEEE Signal Processing Society 17 (11):2029-2039 25. Noor B, Ke C, Haider A, Ghulam M (2012) Coefficient of Variation Based Image Selective Segmentation Model Using Active Contours. East Asian Journal on Applied Mathematics 2 (2):150-169 26. Ali H, Badshah N, Chen K, Khan GA (2016) A variational model with hybrid images data fitting energies for segmentation of images with intensity inhomogeneity. Pattern Recognition 51 (C):27-42 27. Bresson, Esedoglu, Vandergheynst, Thiran, JP, Osher (2007) Fast Global Minimization of the Active Contour/Snake Model. Journal of Mathematical Imaging & Vision 28 (2):151-167 28. Wang XF, Huang DS, Xu H (2010) An efficient local Chan-Vese model for image segmentation. Elsevier Science Inc., 29. Gao X, Wang K, Guo Y, Yang Z, Ma Y (2015) Mass Segmentation in Mammograms Based on the Combination of the Spiking Cortical Model (SCM) and the Improved CV Model. Springer International Publishing, 30. Yi-De MA, Yuan JX, Hong-Juan, Zhang (2012) Self-Adaptive Method Using SCM for Noise Removal in Color Images. Dianzi Keji Daxue Xuebao/journal of the University of Electronic Science & Technology of China 41 (5):754-758 31. Li C, Huang R, Ding Z, Gatenby J, Metaxas D, Gore J (2011) A level set method for image segmentation in the presence of intensity inhomogeneities with application to MRI. IEEE Transactions on Image Processing A Publication of the IEEE Signal Processing Society 20 (7):2007-2016 32. Faisal A, Pluempitiwiriyawej C Active contour using local region-scalable force with expandable kernel. In: IEEE International Conference on Information Science and Technology, Icist, 2012. pp 18-24 33. Chan TF, Vese LA (2001) Active contours without edges. IEEE Press, 34. Guo Yn, Dong M, Yang Z, Gao X, Wang K, Luo C, Ma Y, Zhang J (2016) A new method of detecting micro-calcification clusters in mammograms using contourlet transform and non-linking simplified PCNN. Computer Methods and Programs in Biomedicine 130:31-45 35. de Oliveira Silva LC, Barros AK, Lopes MV (2017) Detecting masses in dense breast using independent component analysis. Artificial Intelligence in Medicine 80:29-38 36. Guo YN, Wang X, Yang Z, Wang D, Ma Y Improved Saliency Detection for Abnormalities in Mammograms. In: Computational Science and Computational Intelligence (CSCI), 2016 International Conference on, 2016. IEEE, pp 786-791 37. Panetta K, Zhou Y, Agaian S, Jia H (2011) Nonlinear Unsharp Masking for Mammogram Enhancement. IEEE Transactions on Information Technology in Biomedicine A Publication of the IEEE Engineering in Medicine & Biology Society 15 (6):918-928 38. Zhan K, Shi J, Wang H, Xie Y, Li Q (2016) Computational Mechanisms of Pulse-Coupled Neural Networks: A Comprehensive Review. Archives of Computational Methods in Engineering:1-16 39. Guo YN, Ma Y (2017) Object detection system based on multimodel saliency maps. Journal of Electronic Imaging 26 (2) 40. Zhan K, Zhang H, Ma Y (2009) New spiking cortical model for invariant texture retrieval and image processing. IEEE Transactions on Neural Networks 20 (12):1980 41. Xu S, Liu H, Song E (2011) Marker-controlled watershed for lesion segmentation in mammograms. Journal of Digital Imaging 24 (5):754 42. Zhang H, Fritts JE, Goldman SA (2008) Image segmentation evaluation: A survey of unsupervised methods. Computer Vision & Image Understanding 110 (2):260-280 43. Guo Yn, Yang Z, Ma Y, Lian J, Zhu L (2017) Saliency Motivated Improved Simplified PCNN Model for object Segmentation. Neurocomputing 44. Chang X, Yu YL, Yang Y, Xing EP (2016) Semantic Pooling for Complex Event Analysis in Untrimmed Videos. IEEE Transactions on Pattern Analysis & Machine Intelligence 39 (8):1617-1632 45. Chang X, Yang Y (2017) Semisupervised Feature Analysis by Mining Correlations Among Multiple Tasks. IEEE Transactions on Neural Networks & Learning Systems 28 (10):2294-2305 46. Li Z, Nie F, Chang X, Yang Y (2017) Beyond Trace Ratio: Weighted Harmonic Mean of Trace Ratios for Multiclass Discriminant Analysis. IEEE Transactions on Knowledge & Data Engineering PP (99):1-1

15

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

47. Lan X, Ma AJ, Yuen PC Multi-cue Visual Tracking Using Robust Feature-Level Fusion Based on Joint Sparse Representation. In: Computer Vision and Pattern Recognition, 2014. pp 1194-1201 48. Lan X, Ma A, Yuen PC, Chellappa R (2015) Joint Sparse Representation and Robust Feature-Level Fusion for Multi-Cue Visual Tracking. IEEE Transactions on Image Processing A Publication of the IEEE Signal Processing Society 24 (12):5826 49. Lan X, Zhang S, Yuen PC Robust joint discriminative feature learning for visual tracking. In: International Joint Conference on Artificial Intelligence, 2016. pp 3403-3410 50. Lan X, Yuen PC, Chellappa R Robust MIL-Based Feature Template Learning for Object Tracking. In: AAAI, 2017. pp 4118-4125

16

Author Biographies

Click here to download Author Biographies Authors Biographies.docx

Ya’nan Guo received the B.E. degree in communication engineering from Lanzhou University, Gansu, China, in 2014. She is currently pursuing the Ph.D. degree in Radio Physics at Lanzhou University mentoring by Dr. Yide Ma. Her current research interests include artificial neural networks, image processing, Wavelet Transform, intelligent algorithms and cognitive visual processing. Yide Ma received the B.S. and M.S. degrees in radio technology from Chengdu University of Engineering Science and Technology, Sichuan, China, in 1984 and 1988, respectively. He received the Ph.D. degree from the Department of Life Science, Lanzhou University, Gansu, China, in 2001.He is currently a Professor in the School of Information Science and Engineering, Lanzhou University. He has published more than 50 papers in major journals and international conferences and several textbooks, including Principle and Application of Pulse Coupled Neural Network (Beijing: Science Press, 2006), and Principle and Application of Microcomputer (Beijing: Science Press, 2006). His current research interests include artificial neural networks, digital image processing, pattern recognition, digital signal processing, and computer vision.

Author Photographs

Ya’nan Guo

Yide Ma